|
1 | 1 | # RapidStamp ring-text handover |
2 | 2 |
|
3 | | -Updated 2026-10-23T07:19:18Z. This file is the operational handover. `docs/data.md` and `docs/research.md` are the Chinese measured notes; `docs/sdk.md` is the published detection plus class-routed CTC / PP-OCRv5 SDK. Those three are not this recognizer. |
| 3 | +Updated 2026-10-23T07:22:09Z. This file is the operational handover. `docs/data.md` and `docs/research.md` are the Chinese measured notes; `docs/sdk.md` is the published detection plus class-routed CTC / PP-OCRv5 SDK. Those three are not this recognizer. |
4 | 4 |
|
5 | 5 | Update this file after every meaningful step, measured result, or change to data or models. Commit it on `main`, then copy `docs/*.md` to `/home/rapidstamp_data/docs/` on the training machine. Do not put passwords, the SSH password, or Aliyun APPCODEs in this file or in git. |
6 | 6 |
|
@@ -55,7 +55,7 @@ A second copy of the collection box is arriving as five plain tar archives, spli |
55 | 55 | | --- | --- | --- | --- | --- | |
56 | 56 | | `a1_meta.tar` | 20 | 442,204,160 | 1,124 | complete. Tar sha256 `f2556116d5aa80cdd48af47887a72a40274fff0d398de67620c711b0da25f0dc`. Extracted. | |
57 | 57 | | `a2_cn_crops.tar` | 60 | 1,361,745,920 | 26,265 | complete. Tar sha256 `b708e5980449559949c136bc1d60127b00d6bbcbb66978b1fcd1218a191fd379`. Extracted into the existing `web_seals_v2_raw/` tree with no file collisions. | |
58 | | -| `a3_foreign_crops.tar` | 26 | 591,646,720 | 4,787 | 000–021, sha256 checked on the server. | |
| 58 | +| `a3_foreign_crops.tar` | 26 | 591,646,720 | 4,787 | 000–023, sha256 checked on the server. | |
59 | 59 | | `a4_cn_rejected.tar` | 13 | 277,166,080 | 4,350 | none | |
60 | 60 | | `a5_foreign_rejected.tar` | 36 | 813,629,440 | 3,637 | none | |
61 | 61 |
|
@@ -123,7 +123,7 @@ For each query it scrapes Bing, Baidu, and 360 image search, crops seals with th |
123 | 123 | | ViT-Tiny direct reader, epoch 33 of 40 | 9/344, 0.31392 | 0/63, 0.18681 (verified 0/57, 0.19851; pending 0/6, 0.07564) | Best by val exact. Predictions are mostly generic shells. One val near-miss: 吉南省人力资源和社会保障厅 vs 吉林省人力资源和社会保障行. | |
124 | 124 | | PARSeq ViT-S pretrain, in progress | see below | see below | Not a result yet. | |
125 | 125 |
|
126 | | -Pretraining snapshot, job still running, log tail at 2026-10-23T07:19:18Z (PID 52706): seen 749,024, loss 4.558, 110.4 img/s. Eval so far (1−NED). Latest val exact is 75/344 at 725,024, up from 68/344 at 675,008, so this row wins `(exact, 1−NED)`. Gold exact is 1/63, the third time (also at 625,024 and 700,000). |
| 126 | +Pretraining snapshot, job still running, log tail at 2026-10-23T07:22:09Z (PID 52706): seen 769,216, loss 4.505, 110.5 img/s. Eval so far (1−NED). Latest val exact is 59/344 at 750,016, down from 75/344 at 725,024, so best stays on 725,024. Gold exact is 0/63. |
127 | 127 |
|
128 | 128 | | Seen | Val 1−NED (n=344) | Gold 1−NED (n=63) | Verified (57) | Pending (6) | |
129 | 129 | | --- | --- | --- | --- | --- | |
@@ -156,8 +156,9 @@ Pretraining snapshot, job still running, log tail at 2026-10-23T07:19:18Z (PID 5 |
156 | 156 | | 675,008 | 0.71306 | 0.59534 | 0.60888 | 0.46666 | |
157 | 157 | | 700,000 | 0.72041 | 0.58637 | 0.59806 | 0.47540 | |
158 | 158 | | 725,024 | 0.72596 | 0.62392 | 0.64079 | 0.46364 | |
| 159 | +| 750,016 | 0.71827 | 0.59805 | 0.61030 | 0.48176 | |
159 | 160 |
|
160 | | -`eval.json` on disk matches the 725,024 row. Val exact is 75/344, so this row wins `(exact, 1−NED)` over 68/344 at 675,008. Gold exact is 1/63. Gold 1−NED rose to 0.62392. Gold exact hits so far are at 625,024, 700,000, and 725,024. Predictions are still mostly generic shells outside those gold hits. Loss started near 10 (two-pass CE; `log(6627) ≈ 8.8`) and is falling. These rows are early. The zero-shot number is the eval written when the 2,000,000-sample stage ends. |
| 161 | +`eval.json` on disk matches the 750,016 row. Val exact is 59/344, below 75/344 at 725,024, so best stays on that earlier checkpoint. Gold exact is 0/63. Gold 1−NED fell to 0.59805. Gold exact hits so far are at 625,024, 700,000, and 725,024. Predictions are still mostly generic shells outside those gold hits. Loss started near 10 (two-pass CE; `log(6627) ≈ 8.8`) and is falling. These rows are early. The zero-shot number is the eval written when the 2,000,000-sample stage ends. |
161 | 162 |
|
162 | 163 | ## Checkpoints and ONNX |
163 | 164 |
|
|
0 commit comments